Axis Robotics Open-Sources Large Franka Arm Simulation Dataset for Physical AI

Dataset Release and Adoption

Axis Robotics published Axis Sim Dataset V1, comprising more than 50,000 human-teleoperated simulation trajectories across 207 manipulation tasks and 60,000+ scene variants on a simulated Franka Research 3 arm. The full dataset, training code, and benchmarks are publicly available on Hugging Face, where it has accumulated 160,000 downloads, making it the most downloaded open-source simulation Franka manipulation dataset on that platform.

Benchmark Performance and Technical Approach

In internal benchmarks, continual pretraining on V1 lifted π0.5 performance and outperformed a volume-matched RoboCasa baseline, with all results open and verifiable. The release represents a departure from the assumption that robot demonstrations must be near-optimal or heavily filtered; instead, Axis is building what it describes as a "compounding data engine" that accepts diverse quality demonstrations and refines them iteratively.

Funding and Broader Vision

Axis Robotics raised $12 million in seed funding led by Hack VC, with participation from Nomad Capital, Pi Network Ventures, 10K Ventures, and unnamed angel investors. The company plans to expand beyond simulation into egocentric real-world capture, humanoid locomotion-manipulation systems, and human-gated DAgger post-training as part of a vertically integrated approach to physical AI development.